Image data retrieval in modern integrated-circuit image sensors is generally executed as a two-phase “correlated double sampling” (CDS) operation with respect to four-transistor (4T) pixels. In the first phase, a floating diffusion node of the pixel being “read out” is reset to an initial condition and then sampled to obtain a measure of background noise. In the second phase, charge integrated within a photodiode component of the pixel (i.e., according to the number of photon strikes during a preceding exposure interval) is transferred to the floating diffusion node which is then sampled a second time to obtain a raw measure of the photodiode charge level. To finalize the pixel read-out, the reset sample is subtracted from the photocharge sample (canceling background noise common to both samples) with the difference supplied to an analog-to-digital converter (ADC) for digitization, thus producing a digital pixel value corresponding to the photocharge level less background noise.
While CDS pixel read-out significantly improves noise performance relative to non-correlated techniques, drawbacks remain. In particular, the integrated circuit body effect yields a photo response non-uniformity (PRNU) across the pixel array and thus a fixed pattern noise (FPN) that is becoming more pronounced as feature geometries shrink. More generally, as pixels shrink, it becomes difficult to reduce input referred noise and readout speed with conventional CDS techniques due to the limited conversion gain attainable to maintain reasonable TG feed-through and adequate output line settling time.